llrCUDA

[Jean Penné]

Hi,

I verified the primality of the last prime announced on the Prime Database Verification Status, using llrCUDA V3.8.1 :

./llrCUDA -d -a7 -oVerbose=1 -oDebug=0 -q"1880370^524288+1"
[Sat Jan 20 18:55:48 2018]
Base factorized as : 2*3^2*5*17*1229
Base prime factor(s) taken : 17, 1229
Starting N-1 prime test of 1880370^524288+1
Using complex rational base DWT and generic reduction, FFT length = 1507328, a = 3
[Mon Jan 22 06:51:05 2018]
3^((N-1)/1229)-1 is coprime to N!
[Mon Jan 22 08:40:50 2018]
3^((N-1)/17)-1 is coprime to N!
1880370^524288+1 is prime! (3289511 decimal digits) Time : 123098.306 sec.

The total time is pretty satisfying, but gcd's calculus by giants code took several hours... I have to write a CUDA gcd code!
Regards,
Jean

[pepi37]

Originally Posted by Jean Penné

Hi,

I verified the primality of the last prime announced on the Prime Database Verification Status, using llrCUDA V3.8.1 :

./llrCUDA -d -a7 -oVerbose=1 -oDebug=0 -q"1880370^524288+1"
[Sat Jan 20 18:55:48 2018]
Base factorized as : 2*3^2*5*17*1229
Base prime factor(s) taken : 17, 1229
Starting N-1 prime test of 1880370^524288+1
Using complex rational base DWT and generic reduction, FFT length = 1507328, a = 3
[Mon Jan 22 06:51:05 2018]
3^((N-1)/1229)-1 is coprime to N!
[Mon Jan 22 08:40:50 2018]
3^((N-1)/17)-1 is coprime to N!
1880370^524288+1 is prime! (3289511 decimal digits) Time : 123098.306 sec.

The total time is pretty satisfying, but gcd's calculus by giants code took several hours... I have to write a CUDA gcd code!
Regards,
Jean

What GPU was in this case?
34 hours is OK time :)

[Jean Penné]

Originally Posted by pepi37

What GPU was in this case?
34 hours is OK time :)

As I wrote in a previous post :

The GPU is EVGA GeForce GTX 1080 FTW HYBRID GAMING 8Go
The CUDA version is 8.0.44

The CPU is Intel Core i7-5930K(3.5GHz, 6 cores)

The CPU system is Ubuntu Linux x86_64

Regards,
Jean

[Honza]

Anybody got a Windows binary?
Willing to try RTX 2080 and how it is doing...

[VBCurtis]

I have llrcuda running on ubuntu 16.04, SB-era Xeon with Quadro 5000 GPU.
When testing with k=443 and exponents around 3.3M, the FFT chosen is about 4 times larger than regular sllr64 (917nnn vs 224k).
I tried another exponent just over 6M, and again FFT is roughly 4x larger (~1.5M vs 400k).
llrcuda uses 100% load on one core, 90+% on GPU, but production is marginally worse than using that one CPU core for sllr64 alone.

If FFT size choice is a bug and can be fixed, it appears my GPU could do work at the rate of 3-4 cores of this system, which would be quite nice!

Previous versions have shown better speed with smaller k; I'll next try k=13 and report any improvement.

[R. Gerbicz]

Originally Posted by Jean Penné

Mon Jan 22 06:51:05 2018]
3^((N-1)/1229)-1 is coprime to N!
[Mon Jan 22 08:40:50 2018]
3^((N-1)/17)-1 is coprime to N!
1880370^524288+1 is prime! (3289511 decimal digits) Time : 123098.306 sec.

The total time is pretty satisfying, but gcd's calculus by giants code took several hours... I have to write a CUDA gcd code!
Regards,
Jean

You can get a trivial speedup here:
gcd(c0,n)=1 && gcd(c1,n)=1 ... gcd(c_t,n)=1 is true iff gcd(r,n)=1, where
r=(c0*c1*c2*...*c_t)%n
so in every case where you'd use the Generalized Pocklington theorem it is enough to compute only one gcd.

[diep]

VB Curtis i'm carefully watching your findings.

The Xeon i do not know how many gflops a single core is on paper double precision, yet if it's a Fermi Quadro 500 GPU it's having 359.04 gflops double precision and 120GB/s bandwidth to the GDDR5 with 152 watt TDP - but that last number is just a Coca Cola Toto number.

What Ghz setting does the Xeon have?

For now blindfolded guess the SB is single core factor 20 slower than the Quadro...

Originally Posted by VBCurtis

I have llrcuda running on ubuntu 16.04, SB-era Xeon with Quadro 5000 GPU.
When testing with k=443 and exponents around 3.3M, the FFT chosen is about 4 times larger than regular sllr64 (917nnn vs 224k).
I tried another exponent just over 6M, and again FFT is roughly 4x larger (~1.5M vs 400k).
llrcuda uses 100% load on one core, 90+% on GPU, but production is marginally worse than using that one CPU core for sllr64 alone.

If FFT size choice is a bug and can be fixed, it appears my GPU could do work at the rate of 3-4 cores of this system, which would be quite nice!

Previous versions have shown better speed with smaller k; I'll next try k=13 and report any improvement.

[VBCurtis]

Originally Posted by diep

What Ghz setting does the Xeon have?

For now blindfolded guess the SB is single core factor 20 slower than the Quadro...

Xeon x5650, 6x2.66ghz. Two CPUs in the box, not that it matters for this llrcuda attempt.

[Stargate38]

Could someone please tell me where to get a Windows 64-bit build?

[Ken_g6]

Originally Posted by Stargate38

Could someone please tell me where to get a Windows 64-bit build?

Last I heard the latest LLR-like prime testing thing for GPUs was Proth20. Binaries here, including Win32. Source and bug tracker here.

It sure looks like llrCUDA hasn't been updated in awhile.

[Stargate38]

What about general numbers of the form k*b^n+/-1? Proth20 only does numbers of the form k*2^n+1, k < 10^8-1). I want to try and find Top 5000 primes with other bases (i.e. b=6, 10, 62, 94, and any other bases that I want to use). Basically, I'm looking for a CUDA equivalent of PFGW.